The Cost of Fairness: Evaluating Economic Implications of Fairness-Aware Machine Learning

With the increasingly popular use of algorithmic-based decision support systems, the issue of discrimination in machine learning has attracted increasing concerns. Essentially, when an algorithm is applied to prediction tasks that are related to human, a prediction algorithm with the objective to solely optimize its predictive power often pick up biases in the training data and use them in the process. Consequently, the prediction results may discriminate one (or more) group of subjects because of the biases captured in the prediction process. In response to this issue, fairness-aware machine learning, which describes machine learning algorithms that can identify biases in the data and suppress their influence in the prediction process, is developed. This paper studies the impact of two most popular types of fairness-aware machine learning, the demographic parity (DP) and the equalized odds (EO), on welfare of stakeholders and society. We demonstrate that the EO is the only type of fairness-aware machine learning that can remove group-level disparity. However, such fairness comes with a non-negligible cost to the firm and society.